There is a sentence couples therapists sometimes say to a couple struggling to communicate: you have to learn to listen to each other, but it does not mean you have to agree. The sentence has to be said because the people in front of them believe the opposite. They believe that listening is a kind of conceding, that to give a real hearing to what their partner is saying is to risk losing ground in an argument neither of them is willing to lose. So they do not listen. They perform listening, while preparing the next position, and the cycle continues until someone hands them a sentence that tells them they have been doing it wrong.

It is a useful sentence in therapy. It is also one of the more accurate descriptions of professional life I have encountered. The conflation between listening and agreeing is not a private problem. It is a defining feature of how people who have to work together fail to do so.

The conflation does its most quiet damage in places where it is least examined: in meetings, in leadership teams, in working groups, in any setting where people are convened on the premise that they are going to think together. These are the settings where listening should be most consequential. They are also the settings where the conflation between listening and agreeing is most powerful, because the cost of agreement, perceived or actual, can be felt in real currency. Status. Position. Whose framing wins. Whose work gets prioritized. People show up to such meetings already calibrated to defend, and they defend by not listening.

What looks like disagreement in those settings is often not. What looks like dialogue is often two or more people speaking past each other from different sides of an observation they share. The cost of this, mostly invisible to the people in the room, is enormous. Not because the meeting goes badly, though it does. Because the actual problem in front of them never gets named.

*

What follows is a composite, drawn from a pattern I have seen across multiple meetings in more than one organization.

In senior-level technology governance meetings of the kind held at large regulated institutions, each line of business is allotted time to present compliance metrics, challenges, and project deliverable updates. Several directors, each a senior voice for their line of business and accountable to their own CIO, would spend their allotted time presenting their PowerPoints, sharing metrics, and firing off their frustrations past each other. One would argue that the framework imposed a uniformity that did not account for the operational resilience realities of their line of business. Others would argue similar points, sometimes conveying that their peers had been treating the same business continuity framework as a compliance exercise rather than a discipline rooted in real resiliency. At one such meeting, one of the directors said, "the artifacts we produce no longer correspond to or represent what our teams are actually doing." Another director agreed. "Exactly," he said. "Which is why we need everyone to take the artifact requirements more seriously." They were diagnosing the same gap from opposite sides: the framework, as administered, was not producing what it claimed to produce. Neither of them heard the other for what they were really saying. They could not, because the meeting had lost its original purpose to surface concerns and had become theatre in front of a governing body that everyone in the room knew was structurally unable to act on, improve, or even stress test their inputs. Under those conditions the only thing left to do was to be on record as having said something defensible. Nothing meaningful was decided. Everyone retrenched to their corner, kept checking their box, and defended their territory, even as they kept misunderstanding the alignment hidden in their own disagreement.

What is striking in that exchange is not that the second director failed to hear his colleague. He heard him. He registered what was said. He nodded. He even used the word exactly. And then he responded to a different sentence than the one that had been spoken. The words landed. The meaning could not be permitted to land. To let the actual claim through, that the framework was not producing what it claimed, would require him to revise something he was responsible for defending. So he heard the words and then, in the same breath, neutralized them by translating them into a familiar framing in which he was not implicated.

Carl Rogers, writing with Richard Farson in 1957, described this exact dynamic. Active listening, as he understood it, required a willingness to enter the other person's frame of reference closely enough that one's own frame of reference might shift. We have, Rogers observed, a natural tendency to judge, to evaluate, to approve or disapprove, and that tendency operates faster than our willingness to be persuaded. The Australian writer Hugh Mackay later called this the courage to listen, which is the right name for it. Listening is not difficult because it requires concentration. It is difficult because it requires a willingness to be moved.

The thing we call listening, in most professional settings, is something else. The systems theorist Otto Scharmer has a name for it: downloading. It is the lowest of the four levels of listening he identifies, and it is the level at which most workplace conversations operate. In downloading, we are not processing what is being said. We are confirming what we already believe, filtering incoming information through the framework we brought into the room, and waiting for evidence that either validates or contradicts our position. The conversation is theatre to us. We are not there to learn. We are there to hold ground.

Edgar Schein, working in the same intellectual neighborhood, named the cultural condition that produces this. Organizations, he observed, are built around telling rather than asking. Leadership is rewarded for advocacy, for clarity, for the confident assertion of positions. It is not rewarded for inquiry, for genuine questions to which the asker does not already have answers. Schein called the alternative humble inquiry, and the descriptor mattered. Inquiry requires humility because real questions involve real uncertainty about the answer, which means the asker might learn something that revises their view. Workplaces built around telling produce people who have not practiced asking. They have not practiced listening either, because the two practices are joined. You cannot do one without the other.

What we encounter in meetings, then, is not a failure of attention. It is the visible surface of a missing discipline. Listening as a practice has not been developed in most of the people doing it, because the institutions that employ them have never asked them to develop it. They have been asked to perform it. That is a different thing.

*

The cost of this is not bad meetings. Bad meetings are the symptom we notice because they are tedious. The actual cost is harder to see. We lose the capacity to recognize alignment when it exists. We lose the ability to solve problems collaboratively, because collaboration depends on being able to hear what is actually being said by the people we are working with. We lose dialogue itself, in the older and more demanding sense of the word. David Bohm, who took the term seriously, distinguished between dialogue and discussion. Discussion shares a root with percussion, the breaking apart of an issue into competing fragments. Dialogue, in Bohm's reading, is something different. The Greek dia means "through," not "two," and logos is "meaning." The image he reached for was of a stream of meaning flowing among and through and between the people in the room. Most of what we call dialogue in organizational life is not that. It is discussion, and most of what we call discussion is parallel monologue.

The therapy sentence is correct. You do have to learn to listen to each other, and it does not mean you have to agree. The harder part of that sentence, the part the therapist does not have to say because the couple in front of her can feel it, is the first half. You have to learn. Listening is a learned discipline. It is not a personality trait, not a sign of empathy, not a competence we are born with. It is a practice. We have not practiced it. That is the work in front of us, individually and collectively, and it is not work most of our organizations are structured to reward. Which means most of the practicing will have to be done in spite of those structures, not because of them.

Two institutions approached AI in opposite directions and missed the same thing.

In the spring of 2025 I was taking the capstone course for my master’s in leadership, and my professors were expressing worry and concern for the future for students.

Their concern was that students would be tempted to cheat more and that they would go out into the world holding credentials that covered vast gaps in their knowledge, that they would lean on an unproven technology, and that they would fail to learn how to think for themselves, becoming dependent on a machine’s inherent bias and false truths.

ChatGPT had exploded into the public social fabric two and a half years earlier. Their concerns had had time to harden into an academic position, with boards still struggling to find an answer to AI slop. The essay had been the academic instrument through which anyone found out whether a person had actually thought about something, and the validity of that instrument had become unreliable. Nobody had a clear answer for an effective replacement. It was just pure worry and resistance to what was changing.

The worry was specific and, I thought, entirely legitimate at the time. I could see what they saw having seen the rapid change that was taking place around AI.

What I could not find in those conversations was the other half of the sentence. Everything named was about loss. Nothing named was about the possibilities. A conversation about disruption that contains only subtraction is not a conversation about disruption. It is a eulogy.

My instincts told me there was more, not just some replacement but a possible magnification of our humanity. Put a capable instrument in the hands of someone educated and driven by curiosity and it does not do their thinking for them. It extends the reach of their mind and accelerates their potential. It lets them get to the third question faster because the first two stopped being expensive. I had no argument with the prediction that people would cheat. Of course they would. It is human nature to seek out efficiency. My objection was that cheating was being treated as the endpoint, and I did not think it was. I’m more interested in the acceleration of curiosity in the hands of the capable.

What made the academic conversation strange was that I was living the opposite of it forty hours a week at work.

The same technology had arrived at my job wearing a completely different costume, and the two arrivals were not sequential. They overlapped. I would spend a week being measured on how aggressively I had adopted these tools and then sit in a classroom being told they were dissolving the foundations of my learning. There was this tension between the demands of my workplace and the academic temple where I was earning my degree.

At first, the adoption at work was silent. Using these tools was something you did quietly, if you did it, and you did not share openly that you had used it to help. Then, quickly enough to give you whiplash, the whole position reversed. Managers now wanted adoption, and they wanted it measured. How fast had you taken it up. How much had you optimized. What was your throughput now compared to before.

This is not unique to any one employer, which is precisely the point. Roughly half of companies now weigh AI usage in performance reviews. Microsoft has told employees that using AI is no longer optional and that managers evaluate them on it. Meta begins weighing AI-driven impact in performance reviews this year. Google and Amazon have issued their own directives. Worker AI use has climbed while workers’ confidence in their own footing has fallen, which is the signature of people complying rather than believing.

What nobody asked, in any meeting I attended, was what the work had become. The question was who could demonstrate more output with less, and everyone in the room understood that to be a question about who would still have a job the following year.

Two institutions, running at the same time, in the same months, reaching opposite conclusions. One of them was certain this would hollow us out. The other was certain it would make us indispensable, provided we moved fast enough. Neither of them was talking about the same thing I was watching, which was what it was doing to the people I knew.

Because underneath the opposition there is a shared assumption, and the assumption is that AI is the variable that determines what happens next.

It isn’t. The variable is the person, and specifically two capacities that we have never been good at measuring and have therefore treated as decorative.

The first is curiosity, by which I mean something narrower than enthusiasm. Curiosity is what produces a question nobody assigned. Without it, a person handed a powerful tool produces faster and more polished versions of what they were already told to make. The output improves. Nothing else does.

The second is empathy, by which I also mean something narrower than kindness. Empathy is the faculty that reads what a situation actually requires, as distinct from what the process says comes next. It is the thing that tells you an answer is technically correct and completely wrong for this moment, this person, this room. It is not a softening agent applied after the analysis. It is part of the perception.

The person who has neither is not a new phenomenon and was not created by any of this. They existed in every organization I have worked in. They learned early to be told what to do and to do that, competently, and to develop a reliable instinct for appearing productive up to the point where appearing productive stops being required. AI does not produce this person. It makes them considerably faster, and speed is easily mistaken for value.

There is a cleaner way to say what separates them from everyone else, and it is not a claim about character.

Problem-solving and problem-finding are different operations. Solving happens inside a problem space that someone else has already defined. It is the part that looks like intelligence because it is the part that has always been graded. But solving for X presupposes that somebody wrote the equation, and writing the equation was never the math.

In the mid-1960s, Jacob Getzels and Mihaly Csikszentmihalyi ran an experiment on art students at the School of the Art Institute of Chicago. Each student was given a table of objects and told to arrange a still life and draw it. What the researchers were actually watching was the part before the drawing. Some students selected a conventional arrangement quickly and got to work. Others handled the objects at length, turned them over, rearranged them, kept deciding what the problem was before committing to an answer. Independent panels judged the second group’s finished work more original. The researchers followed the students long after graduation, and the orientation still tracked with their professional standing. Measured intelligence predicted neither studio performance nor eventual success.

That study is fifty years old. It was conducted on artists, by researchers who had no reason to be defensive about machines, and it says the thing I have been trying to say. The capacity that matters is not the one we have been measuring.

Which brings me to Jensen Huang, who came very close to the same conclusion and then turned left.

Asked on a podcast recently to name the smartest person he has ever met, Huang declined the question. His reasoning was that what we have always called smart, technical problem-solving, is becoming a commodity, and that software programming, long treated as the definitive proof of intelligence, was the first thing AI did well. His own definition of smart is a person at the intersection of technical astuteness and human empathy, able to infer the unspoken, able to preempt a problem before it arrives because they feel the vibe, a vibe he described as assembled from data, first principles, life experience, wisdom, and sensing other people. That person, he added, might score terribly on the SAT.

I have been circling this since 2022 and he says it in ninety seconds. Then comes the last thing he says about them.

Their value is incredible.

And that is where it stops. He has correctly identified what remains once the technical work commoditizes, and he immediately converts it into a statement about which individuals will retain market worth. Nothing in it asks where the capacity comes from. Nothing asks what sustains it, or what destroys it. It is the same sorting logic that produced the adoption dashboards, in a considerably better suit.

Because here is the part that neither institution has been willing to look at. Curiosity and empathy are not endowments. They are conditional, the conditions are not mysterious, and they are trainable and exercisable.

In 1981 Barry Staw, Lance Sandelands and Jane Dutton published what became known as the threat rigidity thesis. Their thesis, built from evidence at the individual, group, and organizational levels, is that perceived threat produces a predictable set of responses: attention narrows, information processing restricts, behavior falls back on well-learned patterns, and control centralizes. Threatened people and threatened organizations stop exploring. They do the thing they already know how to do, just harder. Doubling down.

What makes this worth knowing is that the response is usually correct. Narrowing under threat is an efficient defense when the threat is familiar and the well-learned answer still applies. It fails in one circumstance, which is when the environment has changed enough that the old answer no longer fits, and Staw and his colleagues said so directly. Neither institution I was sitting in was broken. Both were running a sound program against the wrong kind of problem.

Which is also why I want to be careful about the word exploring. What I watched at work was not an absence of motion. It was an enormous amount of motion inside a very narrow channel. Velocity is not exploration. You can move extremely fast in the direction you were already facing.

From the other direction, Karina Schumann, Jamil Zaki and Carol Dweck ran seven studies and found that people who believe empathy is a developable skill rather than a fixed quantity expend more empathic effort precisely where empathy is hardest, listening longer to suffering across racial difference and working harder to understand political opponents. Merely believing the capacity was trainable made people try.

Belief is necessary and it is not sufficient. In a separate line of work, Daryl Cameron, Michael Inzlicht and their colleagues ran eleven studies with more than twelve hundred participants and found that people reliably choose to avoid empathy when given the option, because they experience it as effortful and aversive. Experimentally raising how effective people believed their empathy would be eliminated the avoidance. Which means empathy fails in pressured environments less because anyone thinks it is fixed and more because it costs something they are already spending elsewhere.

Put those together and both institutions indict themselves. An organization that mandates AI adoption under an implicit threat of redundancy is demanding the exact faculty that its own conditions are eliminating. Fear does not manufacture curiosity. Fear forecloses it. You cannot rank people on their velocity and then wonder why nobody is asking interesting questions. And worry is a threat response too. A conversation that stays fixed on what students will get away with has narrowed for the same reason the enterprise narrowed, in the opposite direction, and neither of them ever got around to asking what conditions produce a person who uses this well.

I want to be careful here, because my professors were not wrong about the mechanism.

There is now real evidence for what researchers have started calling metacognitive laziness, the offloading to the machine of the effortful internal work: goal-setting, monitoring your own errors, deciding whether your approach is any good. In one controlled study, students working with generative AI concentrated their cognitive effort on the interaction with the tool rather than on the thinking the tool was supposed to serve. The essays came out better. The knowledge gain did not move.

That is not adjacent to what my professors said. It is the same claim, arrived at from the other direction. They said students would never learn to think for themselves, in the ordinary words available to people who teach, several years before anyone had operationalized it and given it a name. The fear was earned.

They were wrong about the response, but not the risks.

This is the point where someone usually reaches for the industrial revolution, and I want to reach for it too, though not in the direction it normally gets used.

It is offered as reassurance. The machines came, the craftsman was outpaced by mechanical repetition, we adapted, and look at us now. Every clause of that is true and the conclusion is still wrong, because the adapting happened at a scale that contains nobody’s actual life. Between 1780 and 1840, British output per worker rose roughly 46 percent. Real wages rose about 12. The handloom weavers did not retrain into something better. They were immiserated, and the returns arrived two generations later, when output rose 90 percent and wages rose 123. The adaptation was real. It was also paid for in full by specific people who never saw any of it.

I do think this is a moment of evolution rather than a moment of ending, and I do think the doom is half a picture. But optimism is not a prediction. Whether this works out is a question about conditions and distribution, and we have run the experiment once already, under conditions nobody chose on purpose, and we know exactly who absorbed the cost.

What I think is actually happening is that AI does not raise the average. It widens the variance. It amplifies whatever is already present in the person holding it, including the absence of anything.

*

Underneath all of it is attention, which is where this stops being a workplace problem.

In 1971, in an essay called Designing Organizations for an Information-Rich World, Herbert Simon observed that information consumes something, and that what it consumes is the attention of its recipients. A wealth of information, he concluded, creates a poverty of attention. He was writing about organizational design four decades before anyone had a feed, and the observation has since hardened into something closer to a physical law.

Curiosity is how a person decides what is worth their attention. Not how they answer a question. How they choose which question deserves them. That is a finite allocation, made by someone who has to live with the result, and it is the one part of this that cannot be handed off, because handing it off is identical to not making it.

The office is only where I happened to watch it. The same trade is now offered everywhere, several dozen times a day, in terms that are almost always worth taking in the moment. Would you like the answer without the question. Would you like the summary instead of the thing. Would you like to skip the part where you sit with not knowing. Each individual yes is reasonable. It is the aggregate that costs something, and the aggregate is invisible by design.

I have some standing to say that, because I have made the trade myself and I did not hold out.

I started my Master’s degree in late 2019 and finished it in December of 2025, one course a semester, in no particular hurry. When the pandemic moved everything online, I stopped taking classes. My reasoning was that an online-only version of that program would be a thinner version of the thing I had come for, and I would rather wait than take the thinner version. I still think I was right about that, but I also gave in and leaned into the adaptation of the pandemic. I went back, online at first and then hybrid, and it was fine, and I am not sure I would choose differently now.

That is the entire shape of it. The reduced version arrives. You can see clearly that it is reduced. Every reason to accept it is a good reason. You accept it.

*

Here I want to say something I cannot support, and I would rather say plainly that I cannot support it than dress it up.

I do not think these AI systems are independently curious. I think they are extraordinarily good at producing the artifacts that our curiosity produces, and I do not believe anything in them wants to know, like we strive to know. What I often encounter in the output is the shape of an idea that someone prompted. A juxtaposition of human thought and history, composed in new ways, out of material that is entirely ours.

I am aware of what that claim resembles. Tool use was going to be the boundary, and then it wasn’t. Language, theory of mind, art-making, self-recognition, each proposed as the line, each since complicated or crossed. I am making the same shape of argument that has failed repeatedly, and the failure mode is always identical. We mistake the current limits of a technology for the permanent structure of the world, because we have no vantage point outside our own humanity from which to take the measurement. I have no particular reason to think I am exempt.

I hold it anyway, as a hypothesis rather than a finding, and what persuades me has less to do with capability than with wanting. Curiosity, in George Loewenstein’s account, works as a drive state. It is a felt deprivation, produced by noticing a gap in your own understanding, closer in structure to hunger than to preference. It presupposes something with stakes. Something that can be bothered by not knowing.

I know the obvious objection to that, and it is a good one. Machine curiosity is a real research program and has been since Jürgen Schmidhuber proposed, in 1991, rewarding an agent in proportion to its own prediction error. Systems are now routinely built to seek information gain, to chase learning progress, to prefer the regions of a world they model badly. In a precise and functional sense they are drawn toward what they do not yet know. Anyone who tells you these systems cannot be curious has not read the field.

I would still separate the two. An agent minimizing prediction error is optimizing a quantity. Hunger is something that happens to a body that will eventually stop. One is a term in an objective function and the other is a condition of being alive on a deadline. I cannot prove that the difference matters. I have not been able to stop believing that it does.

So the question I would put to a piece of work is not whether it is original. Originality has never been the interesting part, and anyone with an art history education knows how recent and how provincial a value it is. Half of what I love was made by people copying, quoting, and stealing on purpose. The question is whether anything wanted to make it.

Call that a “spark” of humanity if you like. I would rather earn that conception of humanity than presume it is intrinsic.

Which leaves the thing I have no answer to.

If the tool amplifies what is already there, the question was never what AI is going to do to us. It is what conditions we are prepared to protect so that there is something worth amplifying. In something close to four years I have not heard it asked seriously by anyone with the standing to change the answer, in the fearful institutions or the enthusiastic ones.

I no longer think that is an oversight. I think there is a reason we keep talking about the machine instead of our humanity, and it has less to do with what AI can do than with what we believe we are looking at when we look at it.

My thesis on that thought is the next piece I will be writing...